Federated Transfer Learning Strategy: A Novel Cross-Device Fault Diagnosis Method Based on Repaired Data

Area of Science:

• Machine Learning
• Data Science
• Industrial Engineering

Background:

• Federated learning (FL) is crucial for data privacy in fault diagnosis.
• Efficient fault diagnosis requires continuous model training with complete data.
• Data corruption poses significant challenges to FL model training and parameter transfer.

Purpose of the Study:

• To propose a novel cross-device fault diagnosis method using repaired data.
• To address model training difficulties and negative parameter transfer caused by data corruption in FL.
• To enhance fault diagnosis performance while maintaining data privacy.

Main Methods:

• Local model training incorporates random forest regression for repairing data with missing fragments.
• Joint domain discrepancy loss is introduced to mitigate parameter bias during local training.
• An adaptive update strategy is employed for global model aggregation and local model updates.

Main Results:

• The proposed method effectively repairs corrupted fault samples for network training.
• Joint domain discrepancy loss corrects parameter bias, preventing incorrect fault characteristics.
• Adaptive updates manage performance variations from local model updates.
• Experimental validation on bearing datasets demonstrates superior fault diagnosis performance and data privacy protection compared to existing federated transfer learning methods.

Conclusions:

• The novel federated learning approach effectively handles data corruption in fault diagnosis.
• Repaired data and domain discrepancy loss improve model accuracy and robustness.
• Adaptive updates ensure stable and effective federated learning for industrial applications.